Scalable Computing: Practice and Experience, ISSN 1895-1767, http://www.scpe.

org
© 2023 SCPE. Volume 24, Issues 4, pp. 985–997, DOI 10.12694/scpe.v24i4.2475

A DEEP COMMUNITY DETECTION APPROACH IN REAL TIME NETWORKS

DEEPJYOTI CHOUDHURY∗ AND TAPODHIR ACHARJEE†

Abstract. Community detection in real time networks is one of the important aspect of social network analysis. Deep learning
has been applied successfully in a variety of research fields in recent years. Proximity matrix is frequently used as the representation
of the network structure. However, there are issues with the proximity matrix’s insufficient spatial contiguity information. As
a result, this research provides a deep learning applied community identification approach that combines the reorganization of
the matrices, spatial attribute uprooting, and community identification. For obtaining a spatial proximity matrix, the primary
proximity matrices in a real time graph is recreated using the highest weight and adjacent users. The dimensional proximity matrix
can obtain a subdomain of the network, allowing the convolutional neural network (CNN) to draw out dimensional localization
more easily and fast. Ten different real time datasets of social networks are used in tests to examine our proposed approach.
Our results show that the proposed community identification approach has higher compatibility than existing deep learning-based
strategies. As a result, the proposed deep community identification approach is capable of detecting the excellent clusters in real
time networks.

Key words: real-time network, deep learning, community detection, social network, proximity

1. Introduction. It is conventional that social networks have been extensively deliberated to analyze
behaviors of human considering a number of layouts, including information extraction, domination analysis,
community detection, individual profile details, social data privacy etc [1]. Community identification in real-
time social networks is a well-known aspect of networked systems in biology, economics, politics, and computer
science.
Deep learning (DL) has showed excellent performance in a wide number of research domains, including real-
time networks for analysing user structural information [2]. DL-applied network embedding can be executed in
both ways: with random walk [3] and without random walk [4][5]. As we know that proximity matrix is used to
reserve the information of the connected nodes in the network. However, the adjacency matrix has inadequate
spatial proximity information.Several Auto-Encoder (AE)-based network entrenching studies improved the input
vector by performing divergent pre-processing on proximity matrices to improve the elicitation correctness of
spatial characteristics extraction to overcome this complication and boost the correctness of feature uprooting
[6][7][8]. Convolutional neural network (CNN) is used in network embedding because it is an effective technique
for extracting spatial localisation [9][10]. The convolution operation can be simulated on the graph. As a result,
the difficulty is how to enhance the proximity matrix such that it can store spatial closeness among vertices.
1.1. Contribution. Given the resilience and efficacy of AE and CNN-applied network embedding, this
paper integrates ‘AutoEnc+CNN’ to increase the aspect of feature uprooting from the nodes. As a re-
sult, this research offers a deep community detection approach that combines (a) matrix reorganization, (b)
‘AutoEnc+CNN’-applied spatial feature uprooting, and (c) community identification. Furthermore, this paper
provides a spatial characteristics uprooting strategy based on AutoEnc and CNN to uproot the spatial features
of the reorganised proximity matrix.
In recent decades, efforts have been considerably made to build efficient models to identify communities in
social networks. The main objectives of this paper are mentioned below:
1. To obtain the dimensional proximity matrices in real-time networks, a matrix reorganisation strategy
based on a unique structure reorganisation approach is proposed, which can aid CNN in quickly and

∗ Department of Computer Science and Engineering, Assam University, Silchar, Assam, India 788011 (deepjyotichoudhury05@

gmail.com).
† Department of Computer Science and Engineering, Assam University, Silchar, Assam, India 788011 ([email protected]).

985
986 Deepjyoti Choudhury, Tapodhir Acharjee

easily determining geographical localization.

2. To obtain a dimensional features uprooting strategy based on autoencoder and CNN, which is proposed
to derive spatial eigenvectors and successfully uproot the spatial properties of real-time networks.
3. To explore the topographic architecture of real-time networks based on the matrix reorganisation and
dimensional characteristics uprooting approach in order to improve social community detection.
After a brief introduction in Section 1, Section 2 presents important related investigations performed in
recent years. In Section 3, proposed approach is discussed in detailed manner. Execution & Results Analysis
is elaborated in Section 4. We have compared some previous works with our proposed approach in this section.
Section 5 highlights the conclusion, challenges and future work.
2. Related Works. It has been observed in recent years that the applications of Deep Neural Network
(DNN) can embed networks using arbitrary walks while maintaining network properties. DeepWalk is one of
the most well-known deep learning method along with the enhancement of arbitrary walk. Without arbitrary
walk, the same method applies deep approaches over the entire network. AE and CNN are two prominent deep
learning models used in network embedding that do not employ random walks. Resler et al. [11] presented in
their paper that the use of a metric learning enhanced deep CNN on an archaeological dataset. The associations
between sites were determined by them using a community detection algorithm based on the confusion matrix
data. A deep learning-based weighted network community discovery technique was also presented [12] based
on a deep sparse autoencoder. A solution to problem of overlapping community structure for large graph was
presented using autoencoder [13]. Ferraro et al. [14] presented a different technique to resolve community
detection problem based on DL and they proposed a hypergraph-based data model for representing all forms
of user connections within an MSN, which were frequently mediated by multimedia data.
Again, a parallel deep learning-applied community identification strategy using particle swarm optimization
in massive composite networks was proposed [15]. The findings demonstrated how well the suggested deep
learning with hybrid optimisation works for identifying communities in large networks. Essaid et al. [16]
presented a method for detecting communities inside the Bitcoin network that uses a deep feature representation
algorithm and Deep Feedforward Autoencoders. Their findings demonstrated that, compared to a random P2P
network, the Bitcoin network has a stronger clustering coefficient and community structure. Sun et al. [17]
developed a system that combined CNN and Transformer, used wavelet and inverse wavelet transforms for
encoding and decoding, and employed wavelet transform and inverse wavelet transform for learning. A dual
graph autoencoder (DGAE) was suggested by Zhang et al. [18] to develop discriminative representations for
hyperspectral images. In order to characterise the geometric structures of hyperspectral pictures, DGAE first
built the super pixel-based similarity graph with spatial information and band-based similarity graph using the
relationships of pair-wise pixels within homogeneous regions and pair-wise spectral bands. More discriminative
feature representations were learned from the hidden layer via the encoder-decoder structure of DGAE using
the newly created dual graph convolution.
Self-Supervised Contrastive Graph Clustering (SCGC) was suggested by Kulatilleke et al. [19], which
enforced graph structure using contrastive loss signals to acquire discriminative node representations and iter-
atively revised soft cluster labels. A thorough analysis of AE-based industrial applications was published by
Qian et al. [20]. It was primarily divided into two sections: AE-based representation learning and monitoring
techniques, which showed how AE-based monitoring methods are designed from start to finish. Second, a
thorough analysis of AE-based representation learning from the viewpoints of industrial data characteristics
was conducted. A study was carried out by Lim et al. [21] recently in order to offer a thorough overview and to
investigate potential future possibilities for the best reinforcement learning-based virtual network embedding
solutions.
Proximity matrix was well elaborated by Goel et al. [22] and by utilising the user’s own behaviours as
well as those of other users in their social network, they suggested a novel way to build a strong User Interest
Profile (UIP) in community detection. The same authors proposed a methodology [23] that focused on UIP
augmentation using multiple strategies, as well as a novel approach to handle outlier tags that caused ambiguity
in the collective Resource Illustration Profile (RIP). The fuzzy satisfaction requirement-based novel mapping
functions were designed to measure query relevance score and user interest relevance score for a web resource.
A novel approach [24] was proposed that employed Cohen’s k as a similarity measure for each pair of nodes;
 A Deep Community Detection Approach in Real Time Networks 987

the values were then clustered to discover communities. A new community detection approach was presented
by Wu et al. [25] that proved the spatial proximity matrix could obtain a subdomain of the graph, allowing
the convolution neural network to extract spatial localization more easily and fast. An auto-encoder based
on a convolution neural network can extract the spatial eigenvector of the reconstructed adjacency matrix to
improve modularity.
With the aid of some recent works [26][27][28][29][30], we also want to lay the groundwork for how researchers
have implemented deep learning models in the network embedding technique:
• AE-applied Graph Embedding Approach: The AE algorithm is a powerful facts condense tech-
nology. AutoEnc-applied graph embedding approaches frequently modify the parameters of the input
and converts it into a new depiction. To absorb the composition of the network, SDNE was proposed
to design a mini-supervised model based on DNN that enhanced the inputs by combining 1st and 2nd
order adjacency values.
• CNN-applied Graph Embedding Approach: CNN and its variations have found widespread
application in network embedding. CNN-applied graph embedding employs the original CNN model,
which was built for both Euclidean and non-Euclidean domains.
2.1. Motivation. The following are the primary observations that motivated us to present the proposed
work provided in this paper:
1. ‘Matrix reorganisation’ is needed because it is a strategy technique used to improve efficiency, flexi-
bility, and overall performance to identify communities in a large network. The network structure is
restructured from a typical hierarchical model to a matrix form that contains features of both functional
and project-based structures. Matrix reorganisation can convert network data into a more appropriate
structure, such as an adjacency matrix or a modularity matrix, making it easier to discover communi-
ties or groups of nodes within the network. Matrix reorganisation can help increase the scalability of
community detection algorithms in large-scale social networks. It decreases computing complexity and
enables more effective analysis of massive datasets, both of which are critical in comprehending and
managing online communities with millions of members. Matrix reorganization can assist in identifying
influential nodes or key community members within social networks. This is why we have adopted this
approach in our proposed method.
2. ‘Spatial feature uprooting’ method is required because of its ability to select the most discriminating
features, where a deep learning based method can ingest more readily. Communities in many real-
world circumstances are defined not only by social connections but also by geographical proximity.
The inclusion of geographic information into spatial feature uprooting allows community detection
algorithms to incorporate both social and spatial dimensions. Different regions within a network may
have varying degrees of community structure. In order to measure spatial heterogeneity and pinpoint
areas with distinct community boundaries where communities converge, spatial feature uprooting is
used. When spatial linkages and geographic context are important, spatial feature uprooting is crucial
for community detection. It makes it possible to analyse networks more thoroughly by taking into
account both social and spatial dimensions.
3. ‘Community identification’ in real time networks is the final goal based on matrix reorganisation and
spatial feature uprooting method. Communities frequently display recognisable behavioural patterns.
Researchers can learn more about the habits, pursuits, and pursuits of various groups inside the network
by identifying communities. In social networks, community identification is essential for a variety of
purposes, from boosting user experiences and marketing tactics to upholding safety and comprehending
the social dynamics of online communities. It offers insightful information that aids in decision-making
and promotes a better comprehension of the intricate linkages and behaviours seen in social networks.
3. Proposed Approach. As mentioned in the Section 1, our proposed approach is the combination of
matrix reorganization approach, spatial feature uprooting and finally community identification. Let’s elaborate
these approaches in the subsections below.
3.1. Matrix Reorganization Approach. Workflow of the proposed approach has been shown in Fig. 3.1.
The proposed approach comprises of three different sub-approaches: Highest Weight User Selection, Adjacent
988 Deepjyoti Choudhury, Tapodhir Acharjee

Fig. 3.1: Workflow of the Proposed Matrix Reorganization Approach

User Selection and Reorganization of Proximity Matrix. Let’s discuss these sub-approaches below.
3.1.1. Highest Weight User Selection. The user is a key member of a group who may influence the
views of other members using this strategy. In real-time social networks, numerous users can follow and make
friend with the most connected person in a community. As a result, the Highest Weight User Selection technique
proposed here would assess each node’s influence on the next most significant node in the proximity matrix to
identify a starting node for matrix reconstruction. The pseudo codes of this method is presented in Algorithm
1.

Algorithm 1 Proposed Highest Weight User Selection Algorithm

Require: a proximity matrix with nodes and the number of repetitions
Ensure: the user with the highest weight
1: Organize the proximity matrix Anxn
2: while p = 1, ...., n, q = 1, ...., n do
3: Calculate the weight of the connections from user ap to aq based on the number of interconnections of ap
4: end while
5: for p = 1, ...., n do
6: Initialize the weight of user ap in the starting moment as 1.
7: end for
8: while (the amount of repetitions is less than the user having highest weight) do
9: for p = 1, ...., n do
10: Initialize the weight of user ap in the next repetition as 0.
11: for q = 1, ...., n do
12: Compute the weight of user ap grown by the connection weight from user aq to user ap as the merit of
substitute.
13: Initialize the weight of user ap in the next repetition along with the merit of substitute.
14: end for
15: end for
16: end while
17: Return the user has the maximum weight.

3.1.2. Adjacent User Selection. After identifying the person with the highest weight (superior), an
adjacent user selection method is presented to identify the adjacent user who is most relevant to the superior
 A Deep Community Detection Approach in Real Time Networks 989

user.
Based on Equation 3.2, the superior user having the highest weight is represented as user ap . Here, we have
applied Euclidean Distance (ED), z(p, q) to compute the distance between user ap and user aq . According to
Equation 3.3, the user aq with the least distance based on the nearest neighbor is initiated.
v
u n
uX
z(p, q) = t d(p, q, l)2 (3.1)
l=1

where

d(p, q, l) = ap,l − aq,l (3.2)

AdjacentN eighbor = leastz(p,q) (3.3)

where q ̸= p, 1 ≤ q ≤ n
3.1.3. Reorganization of Proximity Matrix. The proposed highest weight user selection and adjacent
user selection methods may be used to determine the order of users in the proximity matrix A. The superior
user may be selected as the starting user, and the adjacent of the superior user can be selected as the next
user; then, using Equations 3.2 and 3.3, the adjacent neighbor of the second user can be selected, and so on.
Algorithm 2 can rebuild the proximity matrix A as matrix Z.

Algorithm 2 Reorganization of Proximity Matrix

Require: a proximity matrix A with n nodes
Ensure: the reorganized proximity matrix Z
1: Organize the proximity matrix Anxn
2: Construct the directory Rn
3: Construct the directory Sn
4: for p = 1, ...., n do
5: Enter the user ap into the directory Rn .
6: end for
7: Initialize the superior leader from the highest weight user selection approach as user ap .
8: while Length of Rn > 0 do
9: Extract user ap from Rn
10: for q = 1, ...., n do
11: Determine the Euclidean distance from user ap to user aq
12: end for
13: Detect as well as initialize adjacent user as user ap
14: Extract user ap from Sn
15: end while
16: Build the proximity y matrix Znxn based on Sn
17: Return the reorganized proximity matrix Z.

3.2. Spatial Feature Uprooting. In a basic instance, both the input as well as output layers have 4
number of neurons, that means 4 users present in the network. The first hidden layer (the convolutional layer)
has a filter size of 1 by 3, therefore two neurons are instructed in the concealed layer. Figure 3.2 depicts the
architecture of a basic CNN-based auto-encoder.
In a normal scenario, the rebuilt proximity matrix may be divided into n records, with each record having
a dimension of 1 by n. Both the input and output layers have n number of neurons. The first concealed layer
has q neurons.
Figure 3.3 depicts the general case structure of a CNN-based auto-encoder. The loss function considers
mean squared deviation (MSD). During the execution and performance stages, spatial information may be
retrieved based on neuron values.
990 Deepjyoti Choudhury, Tapodhir Acharjee

Fig. 3.2: A basic example of a CNN-applied auto-encoder

Fig. 3.3: A general instance of a CNN-applied auto-encoder

3.3. Community Identification. After extracting spatial information, the dimension of each record is
represented as 1xn and used in the K-means method.

Pq = [pq,1 , pq,2 , ...., pq,n ] (3.4)

The proposed community detection approach consists of three phases, which are as follows:
1. The k records are chosen at random from the n records to serve as k cluster centers.

Rp = [rp,1 , rp,2 , ...., rp,n ] (3.5)

2. Based on Equation 3.1, ED is used to calculate the separation space from the q th position to the pth
position of the cluster. Data n are divided into k number of clusters depending on their distance, also
the core of every community is re-evaluated established on the data in the community.
3. If no modifications are made to any of the cluster centers, the community identification process is
completed. Otherwise, Point 1 & 2 must be repeated.
4. Execution & Results Analysis. This section comprises of dataset description, evaluation matrices,
followed by results analysis.
4.1. Dataset. We have considered here 10 different real time datasets to experiment using our proposed
approach. The details is shown in Table 4.1.
4.2. Evaluation Matrices. We have used here Q-modularity, Normalized Mutual Information, Mean
Reciprocal Rank, and Mean Average Precision as the evaluation matrices.
 A Deep Community Detection Approach in Real Time Networks 991

Table 4.1: Dataset [31]

Sl No. Dataset Nodes Edges

1 Karate 34 78
2 Football 115 613
3 Dolphins 62 159
4 Polbooks 105 441
5 Cora 2,708 5,429
6 Facebook 4,039 81,800
7 Artists 50,515 819,306
8 CiteSeer 3,312 4,732
9 Polblogs 1,490 16,718
10 School 68 220

• Q-Modularity [32]: This measure is calculated as:

1 X lp l q
Q= (Apq − )δ(tp , tq ) (4.1)
2m pq 2m

where, A represents an adjacency matrix, m represents the quantity of edge, lp represents the degrees
of the p-th node.
• Normalized Mutual Information [33]: NMI is defined as:
cQ
cP P
P Rlm .N
−2 Rlm log( R l. R.m
)
l=1 m=1
N M I(P, Q) = cP cQ , (4.2)
Rl. log( RNl. ) + R.m log( RN.m )
P P
l=1 m=1

where, cP & cQ are considered as the numeral of the clusters in the partition, P (Q). The total number
of the users in the error matrix are depicted as Rl. (R.m ). N is considered here as the entire users in
the network.
• Mean Reciprocal Rank [34]: MRR is calculated as:
|p|
1 X 1
(4.3)
|p| a=1 qa

where |p| is a total number of queries, qa is the rank position of a first relevant community among all
the communities retrieved for the ath query. A value of MRR ranges between 0 and 1.
• Mean Average Precision: MAP can be defined as:
p
1 X
AvgPa (4.4)
|p| a=1

where |p| is a total number of queries, MAP is the mean of Average Precision (AvgP) of each query in
a query. A value of MAP ranges between 0 and 1.
4.3. Results Analysis. Four scenarios were created to assess the modularity of different approach com-
binations in order to evaluate the proposed approach.
1. To find communities in a social network, the auto-encoder approach is used to extract features from
the original proximity matrix. Case (1)’s label is written as ‘AutoEnc’.
992 Deepjyoti Choudhury, Tapodhir Acharjee

Fig. 4.1: Original Proximity Matrix of Dolphin Network

Fig. 4.2: Reorganized Proximity Matrix of Dolphin Network

2. To discover communities in a social network, the auto-encoder approach is used to extract characteristics
from the rebuilt proximity matrix. Case (2)’s label is written as ‘ReMat+AutoEnc’.
3. To bring out the characteristics of the primary proximity matrix in a real-time network to recognize com-
munity, the CNN-based auto-encoder approach is used. Case (3)’s label is written as ‘CNN+AutoEnc’.
4. To extract the characteristics of a rebuilt proximity matrix in a real-time network to recognize commu-
nity, an CNN-based auto-encoder approach is used. Case (4) is labeled as ‘AutoEnc+ReMat+CNN’.

4.3.1. Execution of Reorganized Proximity Matrices. The process of the reorganization of proximity
matrices is already elaborated in section 3. For the visualisation of the reorganized proximity matrices, we have
considered two datasets as the samples. Figures 4.2 & 4.4 exhibit the reconstructed proximity matrices of
dolphin and karate club network.
 A Deep Community Detection Approach in Real Time Networks 993

Fig. 4.3: Original Proximity Matrix of Karate Network

Fig. 4.4: Reorganized Proximity Matrix of Karate Network

4.3.2. Execution of Evaluation Matrices. This subsection provides the modularity score of all the
four cases as shown in Table 4.2. Among all the four scenarios, ‘AutoEnc+ReMat+CNN’ has attained the
highest modularity score. That is why we have considered the fourth case for comparing the modularity
score with other existing algorithms. Here, we have considered other three popular algorithms for comparison,
namely, ‘Kmeans+NetRA’,‘Kmeans+Node2Vec’, and ‘Kmeans+SDNE’. For all the 10 number of real time
datasets, our proposed approach has gained the highest modularity score for community identification. The
rebuilt proximity matrix is essentially a representation of the network’s structure learned by the auto-encoder.
The quality of the proximity matrix has a significant impact on how accurately communities are detected. A
common technique for dimensionality reduction is auto-encoders. The reconstructed matrix’s properties should
994 Deepjyoti Choudhury, Tapodhir Acharjee

Table 4.2: Modularity Score of 4 Scenarios

AutoEnc+
ReMat+ CNN+
Sl No. Dataset AutoEnc ReMat+
AutoEnc AutoEnc
CNN
1 Karate 0.158 0.272 0.292 0.327
2 Football 0.539 0.629 0.730 0.768
3 Dolphins 0.484 0.539 0.549 0.625
4 Polbooks 0.544 0.593 0.637 0.683
5 Cora 0.358 0.397 0.478 0.483
6 Facebook 0.628 0.728 0.794 0.835
7 Artists 0.472 0.493 0.528 0.632
8 CiteSeer 0.469 0.528 0.573 0.624
9 Polblogs 0.372 0.448 0.472 0.527
10 School 0.573 0.576 0.638 0.735

Table 4.3: Modularity Score Compared with Existing Algorithms

AutoEnc+
Kmeans+ Kmeans+ Kmeans+
Sl No. Dataset ReMat+
NetRA Node2Vec SDNE
CNN
1 Karate 0.273 0.284 0.264 0.327
2 Football 0.528 0.618 0.593 0.768
3 Dolphins 0.528 0.492 0.519 0.625
4 Polbooks 0.492 0.439 0.542 0.683
5 Cora 0.293 0.346 0.274 0.483
6 Facebook 0.639 0.629 0.737 0.835
7 Artists 0.583 0.428 0.484 0.632
8 CiteSeer 0.529 0.553 0.514 0.624
9 Polblogs 0.384 0.418 0.474 0.527
10 School 0.618 0.531 0.683 0.735

preserve pertinent data while becoming less dimensional. Since the network is real-time, the characteristics
of the rebuilt proximity matrix is generated quickly and efficiently. Real-time networks require low-latency
processing, so the auto-encoder approach should be designed to produce the matrix in a timely manner. A
rebuilt proximity matrix generated by a CNN-based auto-encoder is required for accurate, efficient, and flexible
community detection in real-time networks. These qualities influence the approach’s quality of community
detection, scalability, tolerance to noise and changes, and overall efficacy in real-time applications. These
are significance that our proposed approach (Case 4) has improved the performance among all the mentioned
existing methods in essence.
The result is displayed in Table 4.3. Our proposed one, ‘AutoEnc+ReMat+CNN’ has attained the highest
modularity score in Facebook network with 83.5% and achieved lowest in Karate club network with 32.7%.
While ‘Kmeans+NetRA’ generates its better modularity score in Facebook and School network with 63.9% and
61.8% consecutively, ‘Kmeans+Node2Vec’ method provides its best modularity score in Football and Facebook
network with 61.8% and 62.9%. ‘Kmeans+SDNE’ method attained its best modularity score in Facebook
network with 73.7%.
Table 4.4 represents the NMI score comparison with the existing algorithms. In this experiment, our
proposed ‘AutoEnc+ReMat+CNN’ approach has out-beats the other algorithms. Our method has achieved
the highest NMI score in Karate network with 100% follwed by Dolphins network with 95%, Football network
with 93% etc. ‘Kmeans+SDNE’ has also generated the better values of NMI than the other two existing
 A Deep Community Detection Approach in Real Time Networks 995

Table 4.4: NMI Score Compared with Existing Algorithms

AutoEnc+
Kmeans+ Kmeans+ Kmeans+
Sl No. Dataset ReMat+
NetRA Node2Vec SDNE
CNN
1 Karate 0.63 0.81 0.88 1
2 Football 0.59 0.72 0.83 0.93
3 Dolphins 0.65 0.75 0.87 0.95
4 Polbooks 0.62 0.68 0.72 0.78
5 Cora 0.53 0.59 0.60 0.62
6 Facebook 0.70 0.74 0.78 0.84
7 Artists 0.58 0.64 0.68 0.73
8 CiteSeer 0.61 0.67 0.68 0.69
9 Polblogs 0.52 0.63 0.66 0.71
10 School 0.71 0.83 0.85 0.91

Table 4.5: MRR Score Compared with Existing Algorithms

AutoEnc+
Kmeans+ Kmeans+ Kmeans+
Sl No. Dataset ReMAT+
NetRA Node2Vec SDNE
CNN
1 Karate 0.62 0.69 0.78 0.82
2 Football 0.66 0.73 0.81 0.84
3 Dolphins 0.71 0.79 0.83 0.88
4 Polbooks 0.68 0.74 0.79 0.83
5 Cora 0.63 0.71 0.75 0.81
6 Facebook 0.72 0.78 0.84 0.90
7 Artists 0.65 0.74 0.79 0.85
8 CiteSeer 0.63 0.69 0.74 0.80
9 Polblogs 0.59 0.66 0.71 0.78
10 School 0.75 0.80 0.84 0.92

algorithms. Karate club network has achieved 88% followed by Dolphins network with 87%, Footbal network
with 83% NMI score in ‘Kmeans+SDNE’ method. ‘Kmeans+Node2Vec’ method provides its best NMI score
in Karate club network with 81%, followed by Dolphins with 75% and Facebook with 74%. ‘Kmeans+NetRA’
generates its best NMI score in School network with 71%, followed by Facebook with 70%.
MRR score is also considered as one of the evaluation measures which displayed in Table 4.5. After the
experiments, our proposed approach has attained the best MRR score in School network with 92%, followed
by Facebook with 90%. Rest of the datasets have also performed well in our proposed approach. On the other
hand, ‘Kmeans+NetRA’ generates its best MRR score in School network with 75%, followed by Facebook with
72% and Dolphins with 71%. ‘Kmeans+Node2Vec’ method provides its best MRR score in School network with
80%, followed by Dolphins network with 79% and Facebook with 78%. ‘Kmeans+SDNE’ has also generated
the better values of MRR than the other two existing algorithms. It has achieved 84% in both Facebook and
School network, followed by Dolphins with 83% and Football with 81%.
Table 4.6 represents the MAP score comparison with the existing algorithms. Our proposed method has
outperformed among all the existing methods. It has attained the highest MAP score in School network with
88%, followed by Facebook with 87% and Artists with 83%. ‘Kmeans+NetRA’ generates its best MAP score
in School network with 73%. School network has performed well in ‘Kmeans+Node2Vec’ also with 78% MAP
score. But ‘Kmeans+SDNE’ has outperformed the other two methods and generate its best MAP score in
Facebook network with 82%, followed by School with 81%.
996 Deepjyoti Choudhury, Tapodhir Acharjee

Table 4.6: MAP Score Compared with Existing Algorithms

AutoEnc+
Kmeans+ Kmeans+ Kmeans+
Sl No. Dataset ReMAT+
NetRA Node2Vec SDNE
CNN
1 Karate 0.58 0.66 0.76 0.80
2 Football 0.63 0.71 0.78 0.82
3 Dolphins 0.65 0.72 0.75 0.81
4 Polbooks 0.64 0.70 0.76 0.80
5 Cora 0.58 0.68 0.72 0.78
6 Facebook 0.69 0.75 0.82 0.87
7 Artists 0.61 0.71 0.77 0.83
8 CiteSeer 0.60 0.66 0.71 0.78
9 Polblogs 0.57 0.64 0.69 0.75
10 School 0.73 0.78 0.81 0.88

5. Conclusion and Future Work. This research paper proposes a combined auto-encoder and CNN-
based deep community identification approach for real time networks. During our experiments, we have first
evaluated the modularity score on selected datasets using four different cases. Our proposed combined CNN
and auto-encoder based method provides the prominent results on all the datasets. That is why we have
considered our combined approach for evaluation and compare it with other existing approaches. To gather
spatial adjacency matrices and reorganized proximity matrices, a novel matrix reorganization approach is
proposed here. The matrix extends the standard proximity matrix with spatial closeness, obtaining obvious
subspace features, and making convolutional processes simple and rapid to extract network spatial localisation.
In this paper, the ‘AutoEnc+ReMat+CNN’ based approach is designed to obtain spatial eigenvectors, which
spontaneously bring out the graph spatial properties and improve the modularity score. The combined model
of the ‘AutoEnc+ReMat+CNN’based community identification approach serves as the basis for community
identification in a dynamic environment of the network.
The total amount of neurons in the input and output layers remains constant once the DL-based method is
applied in spite of our approach ‘AutoEnc+ReMat+CNN’, which can be a useful investigation of the network
embedding method. The interactions among the users in a real time community may change dynamically.
Therefore, enhanced time-sequence approaches are necessary to draw out spatio-temporal properties for arbi-
trary real-time networks as future approach.

REFERENCES

[1] Yang, D., Liao, X., Wei, J., Chen, G. & Cheng, X. Modeling information diffusion with the external environment in social
networks. Journal Of Internet Technology. 20, 369-377 (2019)
[2] Perozzi, B., Al-Rfou, R. & Skiena, S. Deepwalk: Online learning of social representations. Proceedings Of The 20th ACM
SIGKDD International Conference On Knowledge Discovery And Data Mining. pp. 701-710 (2014)
[3] Mikolov, T., Chen, K., Corrado, G. & Dean, J. Efficient estimation of word representations in vector space. ArXiv Preprint
ArXiv:1301.3781. (2013)
[4] Tao, Z., Liu, H., Li, J., Wang, Z. & Fu, Y. Adversarial graph embedding for ensemble clustering. International Joint
Conferences On Artificial Intelligence Organization. (2019)
[5] Wang, C., Wang, C., Wang, Z., Ye, X., Yu, J. & Wang, B. DeepDirect: Learning directions of social ties with edge-based
network embedding. IEEE Transactions On Knowledge And Data Engineering. 31, 2277-2291 (2018)
[6] Cao, J., Jin, D., Yang, L. & Dang, J. Incorporating network structure with node contents for community detection on large
networks using deep learning. Neurocomputing. 297 pp. 71-81 (2018)
[7] Cavallari, S., Zheng, V., Cai, H., Chang, K. & Cambria, E. Learning community embedding with community detection and
node embedding on graphs. Proceedings Of The 2017 ACM On Conference On Information And Knowledge Management.
pp. 377-386 (2017)
[8] Yu, W., Zheng, C., Cheng, W., Aggarwal, C., Song, D., Zong, B., Chen, H. & Wang, W. Learning deep network representations
with adversarially regularized autoencoders. Proceedings Of The 24th ACM SIGKDD International Conference On
Knowledge Discovery & Data Mining. pp. 2663-2671 (2018)
[9] Xu, Y., Chi, Y. & Tian, Y. Deep convolutional neural networks for feature extraction of images generated from complex
 A Deep Community Detection Approach in Real Time Networks 997

networks topologies. Wireless Personal Communications. 103 pp. 327-338 (2018)

[10] Xu, J., Li, M., Jiang, J., Ge, B. & Cai, M. Early prediction of scientific impact based on multi-bibliographic features and
convolutional neural network. IEEE Access. 7 pp. 92248-92258 (2019)
[11] Resler, A., Yeshurun, R., Natalio, F. & Giryes, R. A deep-learning model for predictive archaeology and archaeological
community detection. Humanities And Social Sciences Communications. 8 (2021)
[12] Li, S., Jiang, L., Wu, X., Han, W., Zhao, D. & Wang, Z. A weighted network community detection algorithm based on deep
learning. Applied Mathematics And Computation. 401 pp. 126012 (2021)
[13] Bhatia, V. & Rani, R. A distributed overlapping community detection model for large graphs using autoencoder. Future
Generation Computer Systems. 94 pp. 16-26 (2019)
[14] Ferraro, A., Moscato, V. & Sperli, G. Deep learning-based community detection approach on multimedia social networks.
Applied Sciences. 11, 11447 (2021)
[15] Al-Andoli, M., Tan, S. & Cheah, W. Distributed parallel deep learning with a hybrid backpropagation-particle swarm opti-
mization for community detection in large complex networks. Information Sciences. 600 pp. 94-117 (2022)
[16] Essaid, M. & Ju, H. Deep Learning-Based Community Detection Approach on Bitcoin Network. Systems. 10, 203 (2022)
[17] Sun, K., Meng, F. & Tian, Y. Multi-level wavelet-based network embedded with edge enhancement information for underwater
image enhancement. Journal Of Marine Science And Engineering. 10, 884 (2022)
[18] Zhang, Y., Wang, Y., Chen, X., Jiang, X. & Zhou, Y. Spectral–spatial feature extraction with dual graph autoencoder
for hyperspectral image clustering. IEEE Transactions On Circuits And Systems For Video Technology. 32, 8500-8511
(2022)
[19] Kulatilleke, G., Portmann, M. & Chandra, S. SCGC: Self-supervised contrastive graph clustering. ArXiv Preprint
ArXiv:2204.12656. (2022)
[20] Qian, J., Song, Z., Yao, Y., Zhu, Z. & Zhang, X. A review on autoencoder based representation learning for fault detection
and diagnosis in industrial processes. Chemometrics And Intelligent Laboratory Systems. pp. 104711 (2022)
[21] Lim, H., Ullah, I., Han, Y. & Kim, S. Reinforcement learning-based virtual network embedding: A comprehensive survey.
ICT Express. (2023)
[22] Goel, S. & Kumar, R. Folksonomy-based user profile enrichment using clustering and community recommended tags in
multiple levels. Neurocomputing. 315 pp. 425-438 (2018)
[23] Goel, S. & Kumar, R. Collaboratively augmented UIP–Filtered RIP with relevancy mapping for personalization of web search.
Information Sciences. 547 pp. 163-186 (2021)
[24] Hoffman, M., Steinley, D., Gates, K., Prinstein, M. & Brusco, M. Detecting clusters/communities in social networks. Multi-
variate Behavioral Research. 53, 57-73 (2018)
[25] Wu, L., Zhang, Q., Chen, C., Guo, K. & Wang, D. Deep learning techniques for community detection in social networks.
IEEE Access. 8 pp. 96016-96026 (2020)
[26] Zhao, S., Peng, R., Hu, P. & Tan, L. Heterogeneous Network Embedding: A Survey. CMES-Computer Modeling In Engi-
neering & Sciences. 137 (2023)
[27] Pham, P., Nguyen, L., Nguyen, N., Kozma, R. & Vo, B. A hierarchical fused fuzzy deep neural network with heterogeneous
network embedding for recommendation. Information Sciences. 620 pp. 105-124 (2023)
[28] Abbas, K., Abbasi, A., Dong, S., Niu, L., Chen, L. & Chen, B. A Novel Temporal Network-Embedding Algorithm for Link
Prediction in Dynamic Networks. Entropy. 25, 257 (2023)
[29] Ding, Y., Zhai, Y., Hu, M. & Zhao, J. Deep Forest Auto-Encoder for Resource-Centric Attributes Graph Embedding. Pattern
Recognition. pp. 109747 (2023)
[30] Vo, T. An integrated topic modeling and auto-encoder for semantic-rich network embedding and news recommendation.
Neural Computing And Applications. pp. 1-16 (2023)
[31] Al-Andoli, M., Tan, S., Cheah, W. & Tan, S. A review on community detection in large complex networks from conventional
to deep learning methods: A call for the use of parallel meta-heuristic algorithms. IEEE Access. 9 pp. 96501-96527 (2021)
[32] Lázár, A., Abel, D. & Vicsek, T. Modularity measure of networks with overlapping communities. Europhysics Letters. 90,
18001 (2010)
[33] Kv, T. On normalized mutual information: measure derivations and properties. Entropy. 19, 631 (2017)
[34] Goel, S., Kumar, R., Kumar, M. & Chopra, V. An efficient page ranking approach based on vector norms using sNorm (p)
algorithm. Information Processing & Management. 56, 1053-1066 (2019)

Edited by: Rajni Mohana

Special issue on: Sentiment Analysis and Affective computing in Multimedia Data on Social Network
Received: Jul 31, 2023
Accepted: Oct 5, 2023